A Big Memory Nvidia GH200 Next to Your Desk: Closer Than You Think

By Doug Eadline

February 22, 2024

Students of the microprocessor may recall that the original 8086/8088 processors did not have floating point units. The motherboard often had an extra socket for an optional 8087 math coprocessor. The math coprocessor made its way into the CPU proper, and today, there are no optional math coprocessors for CPUs.

There are, however, options for SIMD processors, aka GPUs. As we all know, GPUs can speed up mathematical processing (e.g., matrix operations) far beyond that of their CPU hosts.

With the introduction of the Nvidia GH-200 processor  and the AMD MI300A APU, the market is witnessing an “8087 moment” — that is, CPUs absorbing external performance hardware. Both Nvidia and AMD have pulled the GPU into the processor and the result is a big jump in HPC performance with a hint of things to come.

Bye-Bye PCI

GPUs from both AMD and Nvidia rely on the PCI bus to communicate with the CPU. The CPU and GPU have two distinct memory domains, and data must be moved across the PCI interface from the CPU domain to the GPU domain (and back).

A GPU using all 16 lanes in a gen-5 PCIe bus has a maximum bandwidth of roughly 63GB/s. This bottleneck will limit moving memory to and from the CPU and GPU.

The Nvidia GH200 connects the Grace CPU and the Hooper GPU with a 900 GB/s bidirectional NVLink-C2C. That works out to be about 14 times faster. In addition, the GH200 brings the advantage of a single shared CPU-GPU memory domain. Moving data across the PCI bus between CPU and GPU is unnecessary. As shown in Figure 1, the CPU and GPU have a consistent view of all memory. The CPU memory is up to 480GB LPDDR5X (with ECC), and the GPU has either 96GB of HBM3 or 144GB of HBM3e. The total coherent (single domain) memory is between 576GB and 624GB.

Grace Hopper Superchip
Figure 1. Logical overview of the NVIDIA GH200 Grace Hopper Superchip

A single memory domain is presented in the current AMD Instinct MI300A APU with 128 GB of HBM3 memory shared coherently using Infinity Fabric between CPUs and GPUs with 5.3 TB/s on-package peak throughputAlthough the MI300A does not currently support additional DDR memory expansion like the GH200, CXL is one word to remember for the future. 

With both the GH200 and the MI300A, the key standout phrase is “presents a single memory domain.” In the traditional CPU-PCIe-GPU combination, the amount of GPU memory is often less than the CPU memory, and data must be shuffled across the PCIe interface. This bottleneck is eliminated in these two new designs. A single large memory domain has always been attractive to HPC, and the growth of GenAI has accelerated this need (i.e., the ability to load large models in memory and run them using the GPU). With traditional GPUs, the amount of GPU memory has limited the model size and requires a distributed GPU approach. (Note: the GH200 can be connected across the external NVLinks, creating a massive unified memory; e.g., the Nvidia-AWS NLV32 can provide up to 20 TB of unified memory.)

Not that Far Away from Your Desktop

One of the unmistakable trends in technology is the movement from expensive and new to low-cost and commodity markets. HPC is no exception. Everything from multi-core to advanced memory has moved from the top end to “cell phones” as the market demands. The move to a single memory domain is one of these changes.

Recently, on the Linux Benchmark site Phoronix, tester extraordinaire, Michael Larabel, ran HPC benchmarks on a GH200 workstation. The system was made available from GPTshop.ai in Germany.

The system tower chassis, shown in Figure 2, has a GH200 Grace Hopper Superchip with a combined 576G of memory, dual 2000+ W power supplies, a QCT motherboard, and multiple configuration options including SSD and NVIDIA Bluefield/Connect-X adapters. One interesting and useful feature is that the TDP can be programmed from 450W to 1000W (CPU + GPU + memory), which should be useful in non-data center environments. In addition, the default air-cooled noise is reported to be 25 decibels. Liquid cooling is also an option.

The desk-side super workstation is not cheap, however. The currently available model, GH200 576GB, starts at 47,500 € (According to Phoronix, this price translates to $41k due to no 19% VAT when shipped outside the EU)

Figure 2: Internal view of GPTShop Nvidia GH200 Workstation. (Source GPTshop.ai)

 

That price may seem high, but consider that current market prices for a Nvidia H100 PCIe GPU with 80 GB of HBM2e memory run between $30k and $35k. That does not include a host system to power and run the GPU. In addition, users are limited by 80GB of GPU memory that is separated from the main memory domain by the PCIe bus.

The GPTshop workstation offers 576GB of single-domain memory. HPC and GenAI users will find this half TB of CPU-GPU memory appealing.

Preliminary Benchmarks

Phoronix was able to run several benchmarks remotely thanks to GPTshop. The benchmarks should be considered preliminary and not a final performance measure. In particular, the benchmarks were CPU only and did not use the Hopper A100 GPU. Thus, the benchmark picture is incomplete. Phoronix plans to test GPU-based applications in the future.

According to Phoronix, Ubuntu 23.10 with Linux 6.5 was used with GCC-13 as the stock compiler. A similar environment was used to test comparable processors, including Intel Xeon Scalable, AMD EPYC, and Ampere Altra Max processors. A complete list can be found on the Phoronix site.

In addition, there were no power consumption numbers available for the benchmark runs. According to Phoronix, the NVIDIA GH200 doesn’t appear to currently expose any RAPL/PowerCap/HWMON interface under Linux for reading just the GH200 power/energy use. The BMC on the system does expose the overall system power consumption via the web interface, and the power data was not exposed via IPMI.

Despite these limitations, some important benchmarks were run on the GH200 for the first time outside of Nvidia.

Good Ole HPCG

The first test reported by Phoronix was the standard HPCG memory bandwidth benchmark shown in Figure 3.

Figure 3: Results for Nvidia GH200 running HPCG benchmark. (Source Phoronix)

 

As can be seen, the GH200 Arm performance was a respectable 42 GFLOPS and came out just ahead of the Xeon Platinum 8380 2P (40 GFLOPS) and just below the EPYC 9654 Genoa 2P (44 GFLOPS). Also notable is the 72-core Arm Grace CPU, which was nearly twice the performance of the Ampere Altra Max 128-core Arm processor.

The GH200 did well in other benchmarks. The most impressive result is shown in Figure 4. The NWChem (C240-Bucky Ball) run using the 72-core Arm GH200 completed in 1404 seconds, just behind the leader, a 128-core Epyc 9554 (2p) at 1323 seconds.

Figure 4: Results for Nvidia GH200 running NWChem benchmark. (Source Phoronix)

Things to Come

The Nvidia GH200 and the AMD MI300A have introduced a new processor architecture. Similar to the absorption of 8087 math coprocessors, high-end CPUs are beginning to absorb the GPU (or SIMD processing unit). The idea is not entirely new, however. AMD has integrated moderate GPUs as part of their desktop/laptop APU processors since 2011. While these high-end processors may be considered “specialized” and therefore expensive, the huge interest in GenAI may drive these designs into the commodity price point over time. This story will continue to develop as more benchmarks emerge.

In addition, introducing a personal high-performance workstation with enough memory to run some of the largest LLMs next to your desk is an important milestone. Not to mention the ability to run many large memory GPU-optimized HPC applications. Data centers and the cloud will remain the workhorses of the day, but something must be said for “owning the reset button.”

Subscribe to HPCwire's Weekly Update!

Be the most informed person in the room! Stay ahead of the tech trends with industry updates delivered to you every week!

Intel’s Silicon Brain System a Blueprint for Future AI Computing Architectures

April 24, 2024

Intel is releasing a whole arsenal of AI chips and systems hoping something will stick in the market. Its latest entry is a neuromorphic system called Hala Point. The system includes Intel's research chip called Loihi 2, Read more…

Anders Dam Jensen on HPC Sovereignty, Sustainability, and JU Progress

April 23, 2024

The recent 2024 EuroHPC Summit meeting took place in Antwerp, with attendance substantially up since 2023 to 750 participants. HPCwire asked Intersect360 Research senior analyst Steve Conway, who closely tracks HPC, AI, Read more…

AI Saves the Planet this Earth Day

April 22, 2024

Earth Day was originally conceived as a day of reflection. Our planet’s life-sustaining properties are unlike any other celestial body that we’ve observed, and this day of contemplation is meant to provide all of us Read more…

Intel Announces Hala Point – World’s Largest Neuromorphic System for Sustainable AI

April 22, 2024

As we find ourselves on the brink of a technological revolution, the need for efficient and sustainable computing solutions has never been more critical.  A computer system that can mimic the way humans process and s Read more…

Empowering High-Performance Computing for Artificial Intelligence

April 19, 2024

Artificial intelligence (AI) presents some of the most challenging demands in information technology, especially concerning computing power and data movement. As a result of these challenges, high-performance computing Read more…

Kathy Yelick on Post-Exascale Challenges

April 18, 2024

With the exascale era underway, the HPC community is already turning its attention to zettascale computing, the next of the 1,000-fold performance leaps that have occurred about once a decade. With this in mind, the ISC Read more…

Intel’s Silicon Brain System a Blueprint for Future AI Computing Architectures

April 24, 2024

Intel is releasing a whole arsenal of AI chips and systems hoping something will stick in the market. Its latest entry is a neuromorphic system called Hala Poin Read more…

Anders Dam Jensen on HPC Sovereignty, Sustainability, and JU Progress

April 23, 2024

The recent 2024 EuroHPC Summit meeting took place in Antwerp, with attendance substantially up since 2023 to 750 participants. HPCwire asked Intersect360 Resear Read more…

AI Saves the Planet this Earth Day

April 22, 2024

Earth Day was originally conceived as a day of reflection. Our planet’s life-sustaining properties are unlike any other celestial body that we’ve observed, Read more…

Kathy Yelick on Post-Exascale Challenges

April 18, 2024

With the exascale era underway, the HPC community is already turning its attention to zettascale computing, the next of the 1,000-fold performance leaps that ha Read more…

Software Specialist Horizon Quantum to Build First-of-a-Kind Hardware Testbed

April 18, 2024

Horizon Quantum Computing, a Singapore-based quantum software start-up, announced today it would build its own testbed of quantum computers, starting with use o Read more…

MLCommons Launches New AI Safety Benchmark Initiative

April 16, 2024

MLCommons, organizer of the popular MLPerf benchmarking exercises (training and inference), is starting a new effort to benchmark AI Safety, one of the most pre Read more…

Exciting Updates From Stanford HAI’s Seventh Annual AI Index Report

April 15, 2024

As the AI revolution marches on, it is vital to continually reassess how this technology is reshaping our world. To that end, researchers at Stanford’s Instit Read more…

Intel’s Vision Advantage: Chips Are Available Off-the-Shelf

April 11, 2024

The chip market is facing a crisis: chip development is now concentrated in the hands of the few. A confluence of events this week reminded us how few chips Read more…

Nvidia H100: Are 550,000 GPUs Enough for This Year?

August 17, 2023

The GPU Squeeze continues to place a premium on Nvidia H100 GPUs. In a recent Financial Times article, Nvidia reports that it expects to ship 550,000 of its lat Read more…

Synopsys Eats Ansys: Does HPC Get Indigestion?

February 8, 2024

Recently, it was announced that Synopsys is buying HPC tool developer Ansys. Started in Pittsburgh, Pa., in 1970 as Swanson Analysis Systems, Inc. (SASI) by John Swanson (and eventually renamed), Ansys serves the CAE (Computer Aided Engineering)/multiphysics engineering simulation market. Read more…

Intel’s Server and PC Chip Development Will Blur After 2025

January 15, 2024

Intel's dealing with much more than chip rivals breathing down its neck; it is simultaneously integrating a bevy of new technologies such as chiplets, artificia Read more…

Choosing the Right GPU for LLM Inference and Training

December 11, 2023

Accelerating the training and inference processes of deep learning models is crucial for unleashing their true potential and NVIDIA GPUs have emerged as a game- Read more…

Comparing NVIDIA A100 and NVIDIA L40S: Which GPU is Ideal for AI and Graphics-Intensive Workloads?

October 30, 2023

With long lead times for the NVIDIA H100 and A100 GPUs, many organizations are looking at the new NVIDIA L40S GPU, which it’s a new GPU optimized for AI and g Read more…

Baidu Exits Quantum, Closely Following Alibaba’s Earlier Move

January 5, 2024

Reuters reported this week that Baidu, China’s giant e-commerce and services provider, is exiting the quantum computing development arena. Reuters reported � Read more…

Shutterstock 1179408610

Google Addresses the Mysteries of Its Hypercomputer 

December 28, 2023

When Google launched its Hypercomputer earlier this month (December 2023), the first reaction was, "Say what?" It turns out that the Hypercomputer is Google's t Read more…

AMD MI3000A

How AMD May Get Across the CUDA Moat

October 5, 2023

When discussing GenAI, the term "GPU" almost always enters the conversation and the topic often moves toward performance and access. Interestingly, the word "GPU" is assumed to mean "Nvidia" products. (As an aside, the popular Nvidia hardware used in GenAI are not technically... Read more…

Leading Solution Providers

Contributors

Shutterstock 1606064203

Meta’s Zuckerberg Puts Its AI Future in the Hands of 600,000 GPUs

January 25, 2024

In under two minutes, Meta's CEO, Mark Zuckerberg, laid out the company's AI plans, which included a plan to build an artificial intelligence system with the eq Read more…

China Is All In on a RISC-V Future

January 8, 2024

The state of RISC-V in China was discussed in a recent report released by the Jamestown Foundation, a Washington, D.C.-based think tank. The report, entitled "E Read more…

Shutterstock 1285747942

AMD’s Horsepower-packed MI300X GPU Beats Nvidia’s Upcoming H200

December 7, 2023

AMD and Nvidia are locked in an AI performance battle – much like the gaming GPU performance clash the companies have waged for decades. AMD has claimed it Read more…

Nvidia’s New Blackwell GPU Can Train AI Models with Trillions of Parameters

March 18, 2024

Nvidia's latest and fastest GPU, codenamed Blackwell, is here and will underpin the company's AI plans this year. The chip offers performance improvements from Read more…

Eyes on the Quantum Prize – D-Wave Says its Time is Now

January 30, 2024

Early quantum computing pioneer D-Wave again asserted – that at least for D-Wave – the commercial quantum era has begun. Speaking at its first in-person Ana Read more…

GenAI Having Major Impact on Data Culture, Survey Says

February 21, 2024

While 2023 was the year of GenAI, the adoption rates for GenAI did not match expectations. Most organizations are continuing to invest in GenAI but are yet to Read more…

The GenAI Datacenter Squeeze Is Here

February 1, 2024

The immediate effect of the GenAI GPU Squeeze was to reduce availability, either direct purchase or cloud access, increase cost, and push demand through the roof. A secondary issue has been developing over the last several years. Even though your organization secured several racks... Read more…

Intel’s Xeon General Manager Talks about Server Chips 

January 2, 2024

Intel is talking data-center growth and is done digging graves for its dead enterprise products, including GPUs, storage, and networking products, which fell to Read more…

  • arrow
  • Click Here for More Headlines
  • arrow
HPCwire